{"slug": "20-claude-agents-for-m-a-diligence-built-on-one-rule-cite-the-source-or-cut-the", "title": "20 Claude agents for M&A diligence, built on one rule: cite the source or cut the claim", "summary": "A developer built a pack of 20 AI agents for M&A due diligence, called The Deal Team, that enforces a 'cite or cut' rule: every factual claim must trace back to a source document or be discarded. The MIT-licensed, free tool is designed to help small acquirers avoid trusting AI summaries that may contain invented or misleading figures, and it surfaces document contradictions and unverifiable claims as action items.", "body_md": "Most people use AI to summarize a document.\n\nIn an acquisition, the summary is the one thing you can't afford to trust.\n\nIf a model invents an EBITDA figure, or quietly rounds a churn number in the seller's favor, or softens a customer-concentration risk into something\n\nthat *sounds* fine, that isn't a typo you catch later. It's the basis you priced a deal on. By the time you find out, you've signed the LOI, burned a\n\nmonth, and spent credibility you can't get back.\n\nSo I built a pack of agents the other way round. Not one big agent that reads everything and hands you a confident paragraph. Twenty small ones, each\n\nwith a single job, each built to show its work, and to cite every factual claim back to the source it came from, or cut the claim before you ever\n\nsee it.\n\nIt's called **The Deal Team**. It's MIT-licensed, free, and [up on GitHub](https://github.com/MariusGithub13/open-deal-team).\n\n## Why small acquirers feel this hardest\n\nA two-person search fund reviews the same volume of CIMs, financials, and legal documents as a mid-market PE firm. Same teasers, same data rooms,\n\nsame seven-figure decisions, minus the analysts, the associates, and the back office that institutional buyers take for granted.\n\nThat's where AI looks like an obvious win. And it is, right up until you remember that the failure mode isn't \"the summary is a bit rough.\" The\n\nfailure mode is a number that's wrong, stated with total confidence, that you can't tell apart from a number that's right.\n\nReading a CIM, tying it to the financials, and building a question list is days of principal time per target, most of it spent on deals that die. You\n\n*want* to move faster. You just can't move faster by trusting output you can't trace.\n\nThe fix isn't a smarter summarizer. It's a discipline: every figure traces back to a line in a document, or it doesn't get used.\n\n## What \"cite or cut\" actually looks like\n\nThe clearest demonstration isn't in the prompts, it's in a finished brief. There's a [full sample brief](https://os.devaland.com/sample-brief) on a\n\nsynthetic target, \"Project Sentinel,\" a fire-and-safety inspection business. It's worth opening because it shows the discipline doing the one thing a\n\nplain summary can't: catching a document contradicting *itself*.\n\nThe CIM's page 8 says the customer base is well diversified, with no single customer over 15% of revenue. Page 9 of the same document lists the\n\nactual accounts, and the top one runs about 22% of revenue. Both pages can't be true. A model that just summarizes the CIM repeats the page-8\n\ndiversification line, and you price the deal on a concentration risk you never saw. The brief instead shows you both passages, verbatim, each cited\n\nto its page, and leaves the call to you.\n\nThen there's the part most AI tools never show: the discard log. While drafting, the model produced the claim \"the company has no outstanding\n\nlitigation.\" Nothing in the documents supported it, so it was cut before it reached the brief and turned into an action instead: request a 10-year\n\nlitigation history. Same fate for a tidy \"average contract value of about $18,000\" (couldn't be tied to any single sourced passage) and \"technician\n\nretention is strong\" (the CIM said nothing about turnover).\n\nThat's the whole philosophy on one screen. A verified claim shows its source. A contradiction shows both sides. An unverifiable claim doesn't reach\n\nyou at all, it becomes a question.\n\n## What's actually in the pack\n\nTwenty agents, ordered to follow a real deal across five stages. Feed the output of one into the next.\n\nThe full bench, all twenty, each with what it does, is in the [README](https://github.com/MariusGithub13/open-deal-team).\n\n## How to run them (about thirty seconds)\n\nEach agent is a single `SKILL.md`\n\nwith a copy-ready prompt. Two ways in:\n\n**Copy-paste, no setup.** Open claude.ai, copy the prompt from any agent, paste it as your first message, then attach the deal: teaser, CIM, P&L,\n\nyour thesis. For one you'll reuse, drop the prompt into a Claude Project's custom instructions and every chat in that project *is* that agent.\n\n**Install as Claude skills.** Each folder under `skills/`\n\nis self-contained. Drop the ones you want into your skills directory and Claude reaches for\n\nthe right agent when you describe what you're doing:\n\n```\n  git clone https://github.com/MariusGithub13/open-deal-team.git\n  cp -r open-deal-team/skills/* ~/.claude/skills/\n```\n\n## The one habit that makes any of this safe\n\nWhen an agent gives you a number that matters, check the citation it points to. Every figure should trace back to a line in your document. If it\n\ncan't, treat it as unverified.\n\nThat single habit, check the source every time, is the difference between *using* AI on a deal and *trusting* it. The agents are built to make that\n\ncheck fast: they tell you where each claim came from, and they tell you when something they needed was missing instead of papering over the gap.\n\n## From prompts to product, the honest version\n\nI'll be straight about where this ends and where my product begins, because pretending otherwise would undercut the whole \"cite it or cut it\" point.\n\nThe agents give you the discipline: a prompt you run yourself, one document at a time, checking each citation by hand. That's genuinely most of the\n\nvalue, and it's free forever. If you only ever use the prompts, that's a win, they're yours, MIT, no strings.\n\nThe by-hand part is also the part that stops scaling. One agent, one document, one manual citation check is fine for a single target. Across a live\n\ndata room, a CIM, three years of financials, tax returns, a QoE, a stack of contracts, all of which have to agree with each other, the checking *is*\n\nthe work, and it's the work nobody has time for.\n\nThat's what I built ** Deal OS** to do automatically: read a whole data room at once, verify every citation against the\n\nIf you're running one deal, use the prompts. If you're running several and the manual checking has stopped scaling, that's where the product earns\n\nits place. You can [read a full source-cited brief on a sample deal](https://os.devaland.com/sample-brief) before deciding either way.\n\n## One honest note\n\nThese are analytical tools to speed up your own judgment and help you read deals faster. They are not investment, legal, tax, or accounting advice,\n\nand they don't replace your own verification or your advisers. Always check an agent's output against the source documents, which is exactly what the\n\ncited rule is designed to let you do.\n\nIf the \"cite it or cut it\" idea is useful to you, the best thing you can do is [star the repo](https://github.com/MariusGithub13/open-deal-team)\n\nso the next searcher who's drowning in CIMs finds it too. Fork it, adapt the prompts to your own thesis, and tell me which agent earns its keep and\n\nwhich one needs work.\n\n*Compiled by Marius Andronie, Devaland, Deal OS.*", "url": "https://wpnews.pro/news/20-claude-agents-for-m-a-diligence-built-on-one-rule-cite-the-source-or-cut-the", "canonical_source": "https://dev.to/devaland/20-claude-agents-for-ma-diligence-built-on-one-rule-cite-the-source-or-cut-the-claim-2fm6", "published_at": "2026-06-20 08:55:20+00:00", "updated_at": "2026-06-20 09:06:50.117108+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-agents", "ai-tools", "large-language-models", "ai-products"], "entities": ["The Deal Team", "GitHub", "Claude", "Project Sentinel", "MariusGithub13"], "alternates": {"html": "https://wpnews.pro/news/20-claude-agents-for-m-a-diligence-built-on-one-rule-cite-the-source-or-cut-the", "markdown": "https://wpnews.pro/news/20-claude-agents-for-m-a-diligence-built-on-one-rule-cite-the-source-or-cut-the.md", "text": "https://wpnews.pro/news/20-claude-agents-for-m-a-diligence-built-on-one-rule-cite-the-source-or-cut-the.txt", "jsonld": "https://wpnews.pro/news/20-claude-agents-for-m-a-diligence-built-on-one-rule-cite-the-source-or-cut-the.jsonld"}}